audit-general

Audit Excel financial models built by AI tools (Claude, GPT, Copilot) for the specific errors LLMs commonly introduce via openpyxl/pywin32: formula-as-text (string formulas missing =), static snapshots where live formulas should exist, SUM boundary errors, orphaned references, broken cross-sheet links, text-masquerading-as-numbers, date serialisation errors, absolute/relative reference confusion, empty placeholder rows, lookup argument errors, missing error handling, and formatting inconsistencies. Use whenever checking an AI-built model, verifying AI-generated Excel output, or reviewing any workbook produced by an LLM. Trigger on "check AI model", "audit AI-built Excel", "verify openpyxl output", "AI model quality check", "validate AI Excel", "check Claude's model", "GPT built this", or any mention of reviewing an Excel file created by AI.

Evaluate Excel financial model structural integrity, scalability, and architectural best practices. Checks whether the model can be extended without breaking (e.g., adding a new year), identifies fragile architecture prone to breaking upon modification, flags modularity failures where calculations are condensed into single "black box" Mega-Formulas instead of broken into auditable steps, verifies clean separation of inputs, calculations, and outputs, and checks consistent category and time-scale structures across sheets with master date spine alignment. Trigger on phrases like "structural review", "scalability check", "can I extend this model", "architecture audit", "is this model robust", "structural integrity", "inputs and calculations mixed together", or any mention of model structure, scalability, or architectural quality.

Orchestrate a comprehensive Excel financial model audit by coordinating all twelve specialist audit agents (Sentry, Logic, AI, Stylist, Efficiency, Lingo, Architect, Hyperlinks, VBA, Power Query, Navigator, Cartographer) and consolidating their findings into a single professional audit report. Extracts the workbook once, manages the sequence of agent execution, collects findings JSON, deduplicates via script, and produces the final consolidated Audit Report — or hands off to ag-detailed-audit-report for the client document. Use this skill whenever the user wants a full model audit, comprehensive review, complete workbook check, or coordinated multi-agent audit. Trigger on "full audit", "comprehensive review", "audit everything", "complete model check", "run all checks", or any request for an end-to-end model audit.

Scan an entire Excel model, trace every formula dependency chain from input to output, and produce Mermaid flowcharts showing how calculations flow — which cells feed which, which sheets depend on which, and where critical paths run. Traces both formula-layer dependencies (direct cell references) and shadow-layer dependencies (data validations, named ranges, tables, conditional formatting). Produces workbook maps, sheet maps, critical path maps, shadow dependency maps, and a dependency register. Trigger on phrases like "trace dependencies", "map the model", "show calculation flow", "dependency chart", "flowchart", "which cells feed which", "critical path", "shadow dependencies", or any request to visualise model structure.

Produce a comprehensive, reader-friendly Excel model audit report in Markdown, organized by content type rather than by agent. The report walks the reader through the model (summary, flowcharts, formula inventory, formatting inventory, text quality, named ranges, structure, hyperlinks, VBA, Power Query) before presenting all audit findings in a consolidated error section. Section 3 lists every unique R1C1 formula pattern with the A1 cell ranges it covers, highlighting pattern breaks. Use this skill whenever the user wants a detailed audit report, client-ready audit deliverable, comprehensive model review document, or full walkthrough-style audit output. Trigger on "detailed audit report", "client audit report", "full model report", "walkthrough audit", or any request for a structured audit document that separates reference sections from error findings.